Waqar Uddin

The buy button was the easy part

November 16, 2026 (2d ago)3 views

Disclosure: I work at Jazz as Principal Evangelist Cloud & AI. This article is written as personal industry analysis. Views are my own.

On 29 September 2025, OpenAI and Stripe put a checkout inside ChatGPT. Etsy sellers first, then more than a million Shopify merchants, with Walmart in the mix. Shopify's president called it the new frontier of retail.

By March 2026, fewer than 30 of those Shopify merchants had gone live with it. OpenAI retired the feature. Its own explanation was that the checkout "did not offer the level of flexibility that we aspire to provide."

That sentence is the most useful thing anyone has said about agentic commerce this year. It is an admission about where the difficulty sits. Not in the payment. In everything the payment rests on.

What actually shipped

The technical design was good, and it is worth understanding before judging the outcome.

The Agentic Commerce Protocol, ACP, is an open standard from OpenAI and Stripe. Its core primitive is a shared payment token. The buyer's payment provider issues a token that is scoped to one merchant, scoped to an amount, and expires quickly. The agent never touches the stored card. The merchant still decides whether to accept the order, still charges the card, still calculates tax, still ships, still handles the return.

Read the production documentation and the division of labour is explicit. The merchant is the merchant of record. OpenAI and the payment providers are not. The customer sees the shop's name on the statement.

That is the right primitive. It separates two questions that card payments normally collapse into one: who is buying, and what are they allowed to spend.

So the rails were sensible. The merchants did not come.

Why it stalled

The short answer is product data, not payments.

An agent cannot sell what it cannot read. It needs to know that the item is in stock, what it costs today, which variants exist, when it arrives, and what happens if it comes back. Most catalogues were built for a human browsing a rendered page, not for a machine asking a question.

Google said this out loud in March, when it shipped an update to its shopping agent that loads real-time product data. The stated purpose was to prevent "out-of-stocks and pricing errors." Think about that. The failure mode was not a broken payment. It was an agent confidently offering a thing that was not there.

There is a second, quieter reason. A checkout inside a chat window takes the transaction away from the merchant's own site, where the merchant controls the upsell, the account, the loyalty programme and the data. Few retailers were in a hurry to hand that over for a fee.

So the count stayed near 30 while the pipeline was described in millions. Forrester had it at roughly 30 in February. The Information reported the same order of magnitude. Two independent counts, same answer.

Two designs, one fight

Google launched the Universal Commerce Protocol at the National Retail Federation show in January 2026. It was co-developed with Shopify, Etsy, Wayfair, Target and Walmart, and endorsed by more than 20 others, including Adyen, American Express, Best Buy, Flipkart, Macy's, Mastercard, Stripe, the Home Depot, Visa and Zalando.

The difference from ACP is scope. ACP standardises the checkout call. UCP covers discovery, cart, checkout, order management and post-purchase support, and it composes with AP2 for payment authorisation and with MCP for tool access.

The difference in strategy is bigger. ACP put the transaction inside the assistant. UCP makes the assistant a surface over the merchant's own machinery, with the retailer remaining the seller of record and Google Pay handling the money in the first release.

Both can be right at the same time, and both are now in the market. Microsoft Copilot and Shopify's agentic plan use ACP. Google's checkout runs in AI Mode and the Gemini app.

But watch where the money actually moves. Google's Direct Offers lets a retailer put a discount in front of a shopper who has signalled intent to buy. Amazon's Buy for Me lets its agent purchase from other websites. The fight is not over the checkout button. It is over who owns the moment before the button, because that is where the advertising is.

The payments industry stopped waiting

While the assistants argued about protocols, the card networks built the identity layer.

Visa launched Intelligent Commerce Connect in April 2026, a single integration point for onboarding, verifying and issuing tokens to agent platforms. Mastercard's Agent Pay Acceptance Framework lets merchants accept agent-initiated payments under their existing card agreements, and it has been running live trials, including a November booking where an AI travel agent searched, paid for and ticketed a real flight without the customer leaving the conversation. American Express shipped its ACE developer kit in April, and with it a promise to back card member purchases made by registered agents.

EMVCo now has a task force working on how the global specifications support agent payments.

So the plumbing exists. A registered agent can be identified, given a scoped credential, and allowed to transact. That was the part everyone assumed would be hard, and it got solved in about twelve months.

The legal answer arrived, and it is narrow

On 4 August 2026, the Ninth Circuit vacated a preliminary injunction that had kept Perplexity's Comet assistant out of Amazon shopping accounts. The court held that when a user directs the agent, the person at the keyboard is the party accessing the retailer's computers under the Computer Fraud and Abuse Act. Not the company that built the browser.

That is a large practical result. It means a retailer cannot use an anti-hacking statute to shut out an agent it dislikes, as long as a real person pointed it at the site.

It is also much narrower than the headlines suggested. The court said in terms that it was not establishing a new legal regime for agentic AI, and that its holding was limited to the meaning of "access" under that one statute. It said nothing about tort claims, nothing about contract, and nothing about who carries the loss when an agent buys the wrong thing.

So the criminal and anti-hacking question is settled. The money question is not.

Worldpay's read of the current state is the honest one. When an agent authenticates properly and a token is issued correctly, liability generally follows the existing rules for authenticated, tokenised transactions. Past straightforward fraud, the allocation between merchant, issuer and agent platform is still being written. American Express is the exception so far, because it has publicly committed to back purchases made by registered agents.

The lawyers have moved on to the sharper version of the question. It is no longer "was this authorised." It is "can you prove what the buyer authorised." That is an evidence problem, and it lands on whoever keeps the records.

The demand is real, the completion is not

Every forecast in this space is large, and they broadly agree with each other.

Bain puts the US agentic commerce market at $300bn to $500bn by 2030, which is 15% to 25% of e-commerce. Morgan Stanley has a base case of $190bn and a bull case of $385bn, or 10% to 20% of online retail. McKinsey's global figure is $3trn to $5trn.

Now the current state. Bain also reports that 30% to 45% of US consumers already use generative AI for product research and comparison, and that AI and agents influenced $3bn of US Black Friday sales. Worldpay found 45% of consumers ready to let an agent complete a purchase, rising to 54% among Gen Z, while 95% reported at least one concern and half said stronger fraud protection would make them trust it more.

Put those together and the shape is clear. Discovery through an assistant is already ordinary. Completion by an assistant is still rare. The gap between those two facts is the entire opportunity, and it is not a checkout interface problem. It is a data and trust problem.

What to do if you sell

Three things, in order, and none of them is "add a buy button."

First, make the catalogue machine-readable. Stock status, current price, variants, delivery promise, return terms, all of it exposed in a form an agent can query without scraping a page. This is boring work and it is the actual integration.

Second, decide deliberately whether you want agent traffic. An agent that compares on price alone will commoditise anything that is not differentiated. If your margin comes from bundle, service or brand, an agent shopping on a spec sheet is not your channel.

Third, keep the receipt. A record of what the agent was told, what it asked for, what it ordered, and what the buyer approved. If disputes move toward "prove the mandate," the merchant with the log wins arguments that the merchant without one loses by default.

What to do if you build here

I work in a market where most orders still end in cash on delivery, cards are thin, and the storefront is often a messaging thread. The card-network version of agent commerce will arrive late and partially.

Which is why I think the useful thing to build here is not an in-chat card checkout. It is the handoff. An agent assembles the intent, the order is confirmed by a human on a channel they already trust, and the payment happens on a rail that actually clears. The catalogue has to be readable, the authorisation has to be recorded, and the return has to work.

Those three pieces are the same everywhere. What changes is the rail underneath.

Certain, signal, unresolved

Certain, because the record says so: Instant Checkout launched on 29 September 2025 and was retired in March 2026 with fewer than 30 Shopify merchants live. Google launched UCP at NRF in January 2026 with a coalition that includes Walmart, Target, Etsy, Shopify, Visa, Mastercard and Stripe. Visa, Mastercard and American Express all shipped agent payment frameworks in the first half of 2026. The Ninth Circuit vacated the Amazon injunction on 4 August 2026 and limited its holding to the meaning of access under the CFAA. Bain sees 15% to 25% of US e-commerce by 2030.

Signal, and this is my reading: the industry spent a year solving the payment, and the payment turned out to be the solvable part. The two things that decide whether this works are whether a machine can read your catalogue and whether you can prove what the buyer authorised. Both are unglamorous, both are owned by the merchant, and neither shows up in a launch video.

Unresolved: whether any agent platform will carry the loss rather than pass it back to the merchant, whether UCP or ACP becomes the default, and whether markets like mine skip the card rails entirely and build agent commerce on wallets and cash. The tell will be dull. A dispute, a merchant, and a log that settles it.

Related reading: There are three prices for the same token for why the interface that owns the intent gets to set the price, The agent moves onto your desk for what happens when the agent sits on the buyer's machine, and The anchor tenant is the whole deal for the infrastructure underneath all of it.

Sources: Stripe on Instant Checkout and the Agentic Commerce Protocol, 29 September 2025, the ACP specification, OpenAI's agentic commerce production guidance on merchant of record, Fast Company on the retreat and the merchant count, CNBC on OpenAI's pivot and Google's real-time product data update, 20 March 2026, Google's Universal Commerce Protocol announcement, CNBC on the UCP launch, 11 January 2026, Wilson Sonsini on the Ninth Circuit decision, the opinion itself, Worldpay on agentic liability, Lowenstein Sandler on the evidence problem, Bain's 2030 forecast, and Morgan Stanley's outlook.